Claude’s Enterprise Play: Who Controls Your AI Agents?

Executive Summary

• The data: VentureBeat’s Pulse survey of 70 enterprises shows Anthropic’s native orchestration share went from 0% in January to 5.7% in February 2026 — while Claude’s model-layer adoption rocketed from 23.9% to 56.2% between January and March.

• The shift: The enterprise AI race is no longer about which model is best. It is now about who owns the infrastructure layer where your agents execute, store memory, and route decisions — a fight Anthropic, Microsoft, and OpenAI are actively engaged in.

• The surprise: The biggest risk for businesses adopting Claude Managed Agents is not capability. It is that session data, execution graphs, and routing logic are stored inside Anthropic’s own systems — not yours.

For the past two years, the AI conversation in most enterprise boardrooms has sounded roughly the same: which model? GPT-4o or Claude? Gemini or Llama? The assumption was that picking the right model meant winning. That frame is now collapsing — and the data that breaks it arrived in a VentureBeat Pulse survey almost nobody outside of engineering departments is discussing.

Anthropic moved from zero presence in enterprise agent orchestration in January 2026 to 5.7% in February. One month. And Claude’s adoption at the model layer went from 23.9% to 56.2% between January and March. Those are not incremental numbers. That is a breakout curve — and it points directly at the infrastructure layer, not the model itself.

What the Numbers Actually Show — and What They Do Not

To be precise: the 5.7% figure represents four respondents out of 70 in the VB Pulse cohort. Anyone describing that as a decisive market share moment is overselling it. But that is also exactly the wrong way to read this data.

The reason VentureBeat flags it as strategically significant is that it marks the first time Claude usage migrated from the model layer into native orchestration in any enterprise tracking survey. Previously, enterprises used Claude as a model — feeding prompts, collecting outputs — but kept orchestration (routing, memory, tool execution, state management) elsewhere: in LangChain, in Microsoft Copilot Studio, in custom internal pipelines. Something shifted in January 2026.

What shifted is Claude Managed Agents — Anthropic’s bid to become not just the model you call but the runtime where your agents actually live. The platform handles execution graphs, state management, memory, tool routing, and evaluation, all within Anthropic’s managed infrastructure. For an enterprise team that wants to deploy agents without building orchestration plumbing from scratch, it is a genuinely compelling pitch. Microsoft and OpenAI still lead enterprise orchestration by a wide margin. The trajectory of Claude’s model adoption, however, suggests orchestration share will follow.

What ‘Managed Agents’ Does to Your Architecture

Strip away the product language and what Claude Managed Agents means in practice is this: your agents’ session data, execution history, and routing logic are stored in a database managed by Anthropic. The agents run in Anthropic’s runtime. The evaluation framework — the mechanism that decides whether an agent completed its task correctly — lives inside Anthropic’s systems.

That is a fundamentally different relationship than using Claude via API. Via API, you call a model and own everything else: memory, state, orchestration logic, evaluations. Via Claude Managed Agents, you are handing those functions to a single vendor.

For a brand manager or COO, this manifests as a question they may not realise they are being asked: Who controls the operating layer of your AI workforce? Most enterprises have not answered that question consciously. They are adopting tools, not making architecture decisions. That gap is going to become expensive.

Epinium data

Across more than 500 brands Epinium has connected to AI-assisted commerce and marketing workflows over five years, the pattern is strikingly consistent: brands typically arrive with four to seven siloed AI tools already running — copywriting, pricing, catalogue optimisation, ad automation — but no unified orchestration layer connecting them. The control plane problem is not theoretical. It is the single most common bottleneck in our onboarding audits.

Not sure who controls your AI agent stack? Epinium’s Transform practice audits your current AI tools and maps a unified orchestration strategy →

The Lock-In Question Nobody Is Asking Out Loud

Here is the contrarian read on all of this: Anthropic’s $30 billion annual revenue run rate — built on 80x growth in a single quarter, as CEO Dario Amodei disclosed — is precisely what makes the vendor lock-in concern real rather than theoretical. This is not a startup that might disappear. This is a company competing at scale, with strong commercial incentives to deepen its hold on enterprise AI infrastructure over time.

What we are seeing at Epinium is that the most exposed enterprises are mid-market brands that adopted AI quickly and broadly in 2024 and 2025 without a parallel governance decision. They are now discovering that “we use Claude” has quietly become “Claude manages our agents” — and they do not have a clear exit strategy if pricing shifts, API terms change, or a competing model outperforms on their specific tasks.

The alternatives are real. LangChain, LlamaIndex, AutoGen, and CrewAI all offer open orchestration frameworks. Microsoft Azure AI Foundry and Amazon Bedrock Agents offer managed orchestration with different lock-in characteristics. None of them are as frictionless as Claude Managed Agents. All of them preserve more architectural control. For a deeper look at how orchestration architectures compare, Epinium’s Agentic AI Explained guide walks through the options in plain language. And if you want to understand how large enterprises are deploying agents across ERP and operations today, the SAP Autonomous Enterprise analysis is the benchmark to read first.

The VentureBeat data is telling a story that will look obvious in retrospect: the enterprise AI race moved from models to infrastructure in early 2026, and most businesses were not watching. Anthropic’s jump from 0% to 5.7% orchestration adoption in a single month is not a victory lap. It is a signal that the real architectural choices are already being made, often without explicit input from the people who should be making them.

The brands that adapt fastest will be those treating agent orchestration as a governance question, not a technology question. That means asking your AI vendors — today — who owns the control plane, and what it would cost to change the answer.

Ready to audit your AI architecture before it locks you in? Epinium’s Transform practice has guided more than 500 brands through AI adoption decisions, from model selection to orchestration strategy and operational deployment. Discover how Epinium maps your AI stack →

What exactly is an enterprise agent control plane?

The control plane is the layer that sits above individual AI models and manages everything those models do in a production workflow: routing requests, storing context between sessions, executing tools, evaluating outputs, and handling failures. Think of it as the operating system for your AI workforce. Choosing a control plane is a larger, stickier decision than choosing a model — because models can be swapped out, but control planes accumulate dependencies over time.

How does Claude Managed Agents differ from using Claude’s API directly?

With the API, your engineering team builds and owns all orchestration logic — memory, routing, state management, evaluations. It requires real technical resources but keeps every component under your governance. Claude Managed Agents handles all of that for you automatically, at the cost of storing session data and execution logic inside Anthropic’s own infrastructure rather than yours. The tradeoff is deployment speed versus architectural independence.

What is the actual vendor lock-in risk, and when does it become a problem?

The risk is not that Anthropic will behave badly — the company is well-funded and structurally sound. The risk is pricing leverage: once your agents’ state, memory, and execution history reside in Anthropic’s systems, migrating to a different orchestration layer means rebuilding those components from scratch. Enterprises that do not negotiate data portability terms before adoption are most exposed. Raise this in the contracting stage, before any technical integration begins.

Should mid-market brands care about this, or is it mainly a large enterprise problem?

Mid-market brands may be more exposed than large enterprises, which typically have dedicated AI architecture teams actively reviewing these decisions. A brand deploying Claude Managed Agents for e-commerce automation, catalogue management, or ad optimisation is making an infrastructure decision that will compound across 18 to 24 months of product development. The time to design for flexibility is at the start of the deployment, not after three internal product cycles have built dependencies on top of a single vendor’s stack.

What are the practical alternatives if you want to avoid single-vendor orchestration lock-in?

Open-source orchestration frameworks — LangChain, LlamaIndex, AutoGen, and CrewAI — allow you to orchestrate Claude or any other model without placing state management inside Anthropic’s systems. Cloud-native managed options like AWS Bedrock Agents, Azure AI Foundry, and Google Vertex AI Agent Builder offer similar managed convenience with different dependency structures. The most resilient architecture uses a model-agnostic orchestration layer that lets you swap models while keeping execution logic, memory, and evaluations under your own governance.